AUTHORS
Suvrajit Maji, PhD, Jie Liu, MD, PhD2, Kelly Bolton, MD, PhD3, Tatyana Gindin, MD, PhD4, Christopher Y Park, MD,PhD5
AFFILIATIONS
1 NYU Grossman School of Medicine, New York, New York, United States
2 Washington University School of Medicine, St. Louis, Missouri, United States
3 Washington University School of Medicine, St. Louis, Missouri, United States
4 NYU Grossman School of Medicine, New York, New York, United States
5 NYU Grossman School of Medicine, New York, New York, United States
ABSTRACT
Background: Germline pathogenic variants in genes associated with myeloid neoplasms are increasingly recognized through clinical sequencing. However, many identified variants remain classified as variants of uncertain significance (VUS), limiting their clinical utility for risk assessment, diagnosis, family counseling, and donor selection. Conventional variant-effect prediction tools primarily rely on sequence conservation to estimate pathogenicity. More recent AI-based approaches such as AlphaMissense incorporate three-dimensional structural information and improve identification of potentially deleterious variants. Nevertheless, these methods provide limited insight into the molecular mechanisms through which variants contribute to disease predisposition. Many germline variants exert their effects not by destabilizing an entire protein, but by disrupting critical protein–protein or protein–nucleic acid interactions that regulate their function and downstream signaling. Characterizing their impact on molecular interactions offers a biologically informed framework for distinguishing pathogenic from benign variants. We sought to develop a robust workflow that leverages protein structural modeling to predict high-confidence novel protein–protein and protein–nucleic acid complexes on a genome-wide scale.
Methods: By integrating quantitative coevolutionary signatures of interfacial residues with structural feature characterization, PrePVAR (Prediction of Pathogenic Variants) ranks protein–protein interfaces and enables variant-level prediction of pathogenicity. PrePVAR uses HINT (experimentally validated interactomes) and PrePPI (computationally predicted interactions), to identify the target-protein partners. Experimentally solved complexes are obtained from the Protein Data Bank; unsolved protein–protein, protein–nucleic acid, and other biomolecular complexes are modeled with AlphaFold 3 (AF3). Each protein is parsed into domains, increasing the number of pairwise PPIs modeled and evaluated. Given two protein sequences and their domains, predicted complexes are generated, and top-ranking models are evaluated using structural and evolutionary features.
Results: We applied PrePVAR to DNMT3A and 68 confirmed interaction partners, generating 12,997 models, quality filtered to 80 protein domain-domain complexes spanning multiple DNMT3A functional domains. PrePVAR successfully reconstructed solved structures of DNMT3A complex interfaces (e.g., DNMT3B-DNMT3L, DNMT3A-TCL1A) and modeled 35/46 complexes lacking structural experimental data (11 low-confidence). Of 313 unique DNMT3A variants identified at NYU, 233 (74%) mutations were single nucleotide variants (SNVs), of which 168 mapped to the DNMT3A models union. A total of 139 SNVs were mapped to the SAM domain (44 Tier 2, 95 VUS) and 29 SNVs to the ADD domain (8 Tier2, 21 VUS). Of these, 52 (31%) mapped to the interface between DNMT3A and modeled protein partners, 40 (24%) to the surface outside interfaces, and 76 (45%) were buried. 68% of known pathogenic SNVs mapped on the DNMT3A surface were predicted to involve interfacial residues. PrePVAR predicted 51% of VUSs located outside of predicted interacting surfaces and 49% mapped to interfacial residues, the latter likely representing candidate pathogenic variants. To determine if PrePVAR predictions identify clinically meaningful pathogenic variants, we evaluated clinical associations of DNMT3A variants using the UK Biobank dataset, hypothesizing that interfacial variants are more likely to be pathogenic than surface variants. Interface variants were associated with a higher risk of myeloid malignancy compared to DNMT3A surface variants and all CH-associated DNMT3A variants.
Conclusion: PrePVAR is an innovative, structure-centric framework that fills a critical gap in the mechanistic interpretation of variants often missed by conventional sequence- and structure-based tools and can prioritize variants that affect protein-protein and protein-nucleic acid interactions. By integrating structural modeling of protein–protein, protein–peptide, and protein–nucleic acid interactions with evolutionary and biochemical constraints, PrePVAR is uniquely positioned to uncover pathogenic variants driving germline disorders, premalignant states, clonal expansions, and hematologic and non-hematologic cancers.
DISCLOSURES
Suvrajit Maji, PhD: No relevant conflicts of interest to declare.